Randomization of Short-Rate Models, Analytic Pricing and

Flexibility in Controlling Implied Volatilities

LECH A. GRZELAKa,b,∗
a Mathematical Institute, Utrecht University, Utrecht, the Netherlands
b Financial Engineering, Rabobank, Utrecht, the Netherlands

Abstract
We focus on extending existing short-rate models, enabling control of the generated implied
volatility while preserving analyticity. We achieve this goal by applying the Randomized Affine
Diffusion (RAnD) method [11] to the class of short-rate processes under the Heath-Jarrow-
arXiv:2211.05014v1 [q-fin.CP] 9 Nov 2022

Morton framework. Under arbitrage-free conditions, the model parameters can be exogenously
stochastic, thus facilitating additional degrees of freedom that enhance the calibration proce-
dure. We show that with the randomized short-rate models, the shapes of implied volatility can
be controlled and significantly improve the quality of the model calibration, even for standard
1D variants. In particular, we illustrate that randomization applied to the Hull-White model
leads to dynamics of the local volatility type, with the prices for standard volatility-sensitive
derivatives explicitly available. The randomized Hull-White (rHW) model offers an almost
perfect calibration fit to the swaption implied volatilities.
Keywords: Randomization, RAnD Method, Randomized Hull-White Model (rHW),
Stochastic Parameters, Stochastic Collocation, Quantization, Short-Rate Models

1. Introduction

Choosing a pricing model, especially in risk management and for Valuation Adjustments
(xVA), is a balancing act between accuracy and pricing speed. Often, the calculation speed
is critical in determining the model of choice, especially when handling large portfolios. The
modern approach for modelling interest rates is based on the Heath, Jarrow, and Morton
(HJM) [13] framework with an arbitrary term-structure of volatility and covariance of forward
rates across maturities. The framework is generic, and essentially any sensible term structure
dynamics can be modelled by the HJM technique. Unfortunately, the number of models giving
rise to analytical formulae for option pricing or even zero-coupon bonds is minimal. Therefore,
models that can generate realistic implied volatilities while allowing for efficient derivative
pricing are in high demand, especially when the portfolios need to be evaluated multiple times
and for different market scenarios due to regulatory pressure.
A family of the HJM models that has gained particular acceptance from practitioners and
academia is the family of affine models [7]. Over many years, various affine models have
been studied, leading to significant contributions regarding efficient pricing, simulation and
calibration (for the overview, see [4]). In particular, the classic affine short-rate models, like
the Hull-White [14] model, are popularised due to the closed formula for zero-coupon bonds
and semi-analytic swaption pricing. On the downside, they lack sufficient flexibility to calibrate
to market-implied volatilities. Additional model flexibility without compromising numerical
efficiency is highly desired. A straightforward approach to incorporate and control the implied
volatility smile and skew into the short-rate models is to define a stochastic or a local volatility
process, for example, in [6, 8, 9]. Although such models give us a closed-form expression for

∗ Corresponding author at Mathematical Institute, Utrecht University, Utrecht, the Netherlands.

Email address: L.A.Grzelak@uu.nl (LECH A. GRZELAK)
The views expressed in this paper are the author’s personal views and do not necessarily reflect the views
or policies of his current or past employers. The author has no competing interests.

1
the zero-coupon bond, pricing more advanced derivatives like options often involves advanced
numerical techniques like Fourier inversion or numerical integration, which is undesired for the
pricing of high-volume derivatives. A survey of the most popular stochastic volatility models
in the interest rate world can be found in [1].
This article focuses on extending existing interest rate models with model parameters to
be random via the so-called randomization method (RAnD) [11]. Randomization by making
the model parameters stochastic is a technique for more flexibility in the stochastic model-
thus, improving the calibration quality while preserving the analytic properties of the models
considered. The concept of randomization represents the uncertainty of potentially hidden
states that are not sufficiently captured by deterministic parameters. Conventionally, under
the affine framework, an extension by a stochastic parameter would require the model to meet
linearity conditions - this does not have to be the case under RAnD. The method requires
an affinity to hold, given a specific realization of the stochastic parameter. In other words,
it builds an outer layer over the affine models and allows a stochastic parameter setting in
that layer. The quadrature rule handles the numerical complexity associated with an infinite
number of parameter values and reduces it to only a few critical parameter realizations. The
selection of these points is based on the moments of the stochastic parameter.
The RAnD method is generic and can be applied to any pricing model. For the sake of
simplicity, however, we focus on one of the most popular interest rate models, namely the
Hull-White (HW) model [14]. In particular, we will show that applying the RAnD method to
the HW model results in a local-volatility type of dynamics. The local volatility dynamics may
seem like a “deal-breaker”; typically, the local volatility models, especially in interest rates,
are notoriously tricky to operate, i.e., closed-form option pricing is hardly possible. However,
although the randomized Hull-White (rHW) model is of the local volatility type, the option
pricing is as complex as the standard, unrandomized model. Furthermore, the presented model
offers fast and accurate calibration as the local volatility dynamics provide more flexibility in
generating volatility smiles and skews for various market conditions.
Our method also addresses a challenging problem stated by Brigo and Mercurio in [3], where
the mixture of lognormals was studied. In their paper, it was shown that great flexibility to
standard models like Black-Scholes was introduced by considering a convex combination of
the associated probability density functions: ω1 f1 (x; σ1 ) + · · · + ωN fN (x; σN ) with ω1 + · · · +
ωN = 1 and some σi > 0, i = 1, . . . , N. However, the problem encountered was the large
number of the model parameters that are difficult to interpret and relate to the corresponding
implied volatilities. As stated in their work: “the absence of bounds on the parameter N
implies that a virtually unlimited number of parameters can be introduced in the dynamics
so as to be used for a better calibration to market data. (...), one has to find the correct
tradeoff between model flexibility and number of parameters so as to avoid both poor calibration
and over-parametrization issues”. However, this problem is resolved with the RAnD method.
The method transitions from a continuous randomizer ϑ to a unique sequence of weights ωi
and associated values σi . In this setting, the model calibration is performed by varying the
parameters of the randomizer, ϑ; therefore, the number of free parameters is manageable.
Moreover, as presented in this work, the varying parameters of ϑ enable a transparent control
of the implied volatilities.
This article comprises five sections. First, the details of the HJM framework and pricing
equations using the RAnD method are provided in Section 2, and a discussion on the arbitrage-
free conditions is covered in Section 2.1. In Section 2.2, we consider the rHW model and derive
the dynamics of the randomized stochastic processes. Section 2.3 discusses different variants
of the rHW model parameter choices, namely the randomization using univariate or bivariate
distributions. Section 3 focuses on derivative pricing; in particular, the pricing equations for
options on zero-coupon bonds and swaptions are provided. Numerical results are presented
in Section 4, where in Section 4.1, the swaption implied volatility surfaces for HW and rHW
models are compared. Section 4.2 analyzes the randomizers’ impact on controlling implied
volatilities for swaptions. Next, the calibration results are presented in Section 4.3, and in
Section 4.4, the extension to the bivariate randomization case is established. Finally, the
convergence results are illustrated in Section 4.5 and Section 5 concludes.

2
2. The Heath-Jarrow-Morton models, affinity and randomization

Here, we will give an informal presentation of the Heath-Jarrow-Morton (HJM) frame-

work and its relation to arbitrage-free short-rate models. We define the probability space
(Ω, F(t), Q) on which the HJM arbitrage-free condition requires that the instantaneous for-
ward rates, f (t, T ), are modelled by the following arbitrage-free SDE:
Z T
df (t, T ) = α(t, T )dt + γ(t, T )dW (t), α(t, T ) = γ(t, T ) γ(t, z)dz. (2.1)
t

The instantaneous forward rate, f (t, T ), is determined by the volatility driver γ(t, T ), which
can be defined as a deterministic function, stochastic process or by a random variable. The
case where γ(t, T ) is defined as a random variable, thus time-invariant, will be associated with
parameter randomization.
We focus here on the short-rate dynamics under the HJM framework. Suppose that f (0, t),
RT ∂
α(t, T ) and γ(t, T ) are differentiable in their second argument, with 0 ∂t f (0, t) < ∞, then
the short-rate process under the HJM framework is given by,

dr(t) = ζ(t)dt + γ(t, t)dW (t), (2.2)

with Z t Z t
∂ ∂ ∂
ζ(t) = α(t, t) + f (0, t) + α(z, t)dz + γ(z, t)dW (z).
∂t 0 ∂t 0 ∂t
A particular class of popular models are the models belonging to the Affine-Diffusion interest
rate models. It is shown [7] that in this class, for process r(t), the discounted characteristic
function is of the following form:
" !#
Z T
φr (u; t, T ) = Et exp − r(s)ds + iur(T ) = exp (A(u; τ ) + B(u; τ )r(t)) , (2.3)
t

with the expectation under risk-neutral measure Q for τ = T − t, where A(u; τ ) and B(u; τ )
form the fundamental solution to the corresponding pricing PDE and satisfying the complex-
valued ODEs, see the work by Duffie-Pan-Singleton [7]. The affinity conditions are rather strict
as they impose a linear structure on the model, i.e., the model dynamics and its covariance
structure needs to be linear in the state variables. The representation in (2.3) is of particular
importance as it gives an explicit closed-form expression for the integrated short-rate process
and therefore determines the price of a zero-coupon bond (ZCB). Models not leading to an
analytic ZCB price are rarely used for pricing purposes. Unfortunately, because of the affinity
constraints, the model extensions for some model parameters to be stochastic are minimal. In
this article, we focus on relaxing this constraint and letting the model parameters be random
while benefiting from the closed-form solutions for the ZCBs and pricing, as presented in (2.3).
We consider a vector Θ = [ϑ1 , . . . , ϑd ]T , with d ∈ N representing the number of randomized
parameters, where each ϑi is a possibly correlated, time-invariant, random variable. A realiza-
tion of ϑi is indicated by θi , ϑi (ω) = θi , and consequently the realization for Θ is indicated by
θ = [θ1 , . . . , θd ]T . Under the RAnD framework, parameters are not driven by a stochastic pro-
cess but only by a random variable; however, an extension to piece-wise stochastic parameters
is also possible (see [11] for details).
In a nutshell, the RAnD method relies on quadrature integration and conditional expec-
tation. Therefore, the technique can be applied wherever we deal with an expectation. In
principle, randomization can be applied to Equation (2.3), which would provide us with a
randomized ZCB; on the other hand, the randomization can be applied at the valuation level,
where the expectation of the expected payoff is computed. We prefer the latter approach to
construct the randomized variant of any interest rate model by simply evaluating the model
on a particular realization of the model parameters. This will be particularly important when
pricing non-standard derivatives.
In the randomized HJM framework, the volatility in (2.1) may be dependent on random
parameters ϑ1 , . . . , ϑd , i.e., γ(t, T ; {ϑ1 , . . . , ϑd }). Let us now consider a stochastic model un-
der the HJM framework for derivative pricing with the price denoted as V (t, r(t; θ)) with θ
3
indicating a vector-valued realization of the model parameter, Θ. V (t, r(t; θ)) may correspond
to a model of Vašiček, Hull-White, Black-Karasiński or any other model that belongs to the
class of HJM [4]. Then, the pricing under randomized parameter space will be given by the
following relation:
Z
V (t, r(t; Θ)) := Et [V (t, r(t; Θ = θ))] = V (t, r(t; Θ = θ))fΘ (θ)dθ, (2.4)
Rd

where the integration takes place over the parameter space Θ ∈ Rd and fΘ (θ) corresponds
to the associated d−dimensional probability density function. Equation (2.4) shows that a
random model parameter can be considered averaging over possible parameter realizations.
From a numerical perspective, the integration in (2.4) is expensive.
The randomization enables for randomizing d−dimensional space of parameters; however,
for simplicity, we will mainly focus on single or bivariate parameter randomization, ϑ1 (ω) = θ1
and ϑ2 (ω) = θ2 , thus also letting for the correlation between stochastic parameters.
Our aim is twofold: we aim to provide numerically efficient methods for the computation
of the pricing Equation in (2.4) and to determine the stochastic differential equation for the
randomized short-rate processes, r(t; ϑ) where ϑ is the chosen random parameter. The insight
into the dynamics of the randomized stochastic process will help us categorise the randomized
processes and the structure of their drift and volatility coefficients.
Theorem 2.1 (Pricing formula for the randomized model). Consider a random variable ϑ,
defined on some finite domain Dϑ := [a, b], with its PDF, fϑ (x), CDF, Fϑ (x) and a realization
θ, ϑ(ω) = θ such that for some N ∈ N the moments are finite, E[ϑ2N ] < ∞. Let V (t, r(t; ϑ))
be the price of a financial derivative depending on the short-rate r(t; θ) with parameter θ, then
the randomized price, V (t, r(t; ϑ)) is given by:
Z N
X
V (t, r(t; ϑ)) = V (t, r(t; θ))dFϑ (θ) = ωn V (t, r(t; θn )) + N , (2.5)
[a,b] n=1

where the pairs {ωn , θn }N

n=1 are the Gauss-quadrature weights and the nodes based on the
parameter cumulative distribution function, Fϑ (·), determined by ζ(ϑ) : R → {ωn , θn }N
n=1
defined in Appendix A.2 and where the error N is defined as:

1 ∂ 2N
N = V (t; r(t, ϑ = ξ)), a < ξ < b. (2.6)
(2N )! ∂ξ 2N

Proof. Define a payoff depending on the short-rate r(t), 0 < t < T , by χ(r, T ), then, for a
stochastic parameter ϑ with the help of conditional expectation, we find:
  
M (t)
V (t, r(t; ϑ)) = Et Et χ(r, T ) ϑ = θ =: Et [V (t, r(t; θ))] , (2.7)
M (T )

with M (t) being the money-savings account. By application of the quadrature rule, Equa-
tion (2.7) becomes:
Z N
X
V (t, r(t; ϑ)) = V (t, r(t; θ))dFϑ (θ) = ωn V (t, r(t; θn )) + N .
[a,b] n=1

Here, the error is given by N , for a < ξ < b, and the pair {ωn , θn }N
n=1 forms the Gauss-
quadrature weights and the nodes based on the parameter distribution, as proven in [5] (p.180,
Theorem 3.6.24). 
Theorem 2.1 gives us an explicit relation between a continuous randomizer ϑ and the
discretization in terms of pairs {wn , θn }N
n=1 , leading to a simplification of the integral in (2.4).
In this paper, these pairs are based on moments of the randomizer, ϑ, and, for completeness,
the detailed computation procedure is given in Appendix A.2.
The associated error N in (2.6) can be interpreted as a cost when moving from the contin-
uous to a discretized random parameter, ϑ → {ωn , θn }N n=1 . To some extent, the pricing with

4
a discrete set of parameter realizations, θn , and the associated probabilities, wn , resembles
the so-called regime-switch method where a finite set of states is defined. However, from a
practical perspective, it is challenging to deal with N pairs of the possible variable states and
the associated probabilities, especially in the calibration procedure.
The procedure described in Appendix A.2, however, simplifies this problem as we can
control the quadrature pairs employing the parameters of the randomizer ϑ. To illustrate this
process, let us consider a stochastic parameter with a randomizer ϑ ∼ N (µ, σ 2 ); thus, the
stochastic parameter, ϑ, is driven by two parameters µ and σ 2 . By application of Appendix
A.2, for any set of parameters µ and σ 2 , we compute the corresponding pairs {ωn , θn }N n=1 ;
therefore, in the calibration procedure we will only vary µ, and σ 2 . This procedure drastically
reduces the number of associated model parameters, thus facilitating the calibration. We also
stress that the set of optimal pairs is solely obtained based on the moments of the randomizing
random variable, ϑ, implying that every random variable with, preferably, closed-form moments
may be used for randomization. In Table A.2, a few selected random variables and their
moments are tabulated. In Section 4.3, we will also classify the randomizers, ϑ, for which the
computations can be greatly reduced or even tabulated.
To emphasize the flexibility of the RAnD method, we focus on the application in the Hull-
White model, where we will apply randomization to each model parameter and consider an
extension to a bivariate case. In the following sections, we consider a finite number of realiza-
tions of ϑ. We will denote them as θ1 , . . . , θN , for some N ∈ N. These “specific” realizations we
will interchangeably call either “collocation” [12] or “quadrature” points. Finally, by ϑ(â, b̂),
we denote that ϑ is a random variable driven by parameters: â, b̂.

2.1. Arbitrage-free conditions under the RAnD method

Before we analyze the specifics of the randomized Hull-White (rHW) model, let us review
the implications of randomization on the pricing, under the affine-diffusion framework, of ZCB
and discuss arbitrage-free aspects of the RAnD method. By setting u = 0 in (2.3), the risk-free
pricing formula for a randomized ZCB, P (t, T ; ϑ) is given by:
Z Z
P (t, T ; ϑ) = P (t, T ; θ)dFϑ (θ) = eA(τ ;θ)+B(τ ;θ)r(t;θ) dFϑ (θ), a < b, (2.8)
[a,b] [a,b]

with A(τ ; θ) := A(0, τ ; θ) and B(τ ; θ) := A(0, τ ; θ) in (2.3) for τ = T − t, and r(t; θ) indicates
a short-rate model with constant parameter realizations θ. By application of Theorem 2.1, the
randomized ZCB is known explicitly, and it is presented in Corollary 2.1.
Corollary 2.1 (ZCBs under Randomized Affine Jump Diffusion Processes). Let r(t; θ) rep-
resent an affine short-rate process with some constant parameter θ. Assuming that the corre-
sponding ChF, φr(T )|ϑ=θ (·), is well defined and 2N times differentiable w.r.t. θ, the randomized
ZCB is given by,
N
X N
X
P (t, T ; ϑ) = ωn P (t, T ; θn ) + N = ωn eA(τ ;θn )+B(τ ;θn )r(t;θn ) + N
n=1 n=1
=: P (t, T ; {θn }N
n=1 ) + N , (2.9)

with A(τ ; θn ), B(τ ; θn ) being the real-valued functions obtained from Riccati-type of ODEs
available for affine models [16].
Proof. The proof is a consequence of Theorem 2.1. 
Representation (2.9) illustrates that the ZCB, P (t, T ; ϑ), randomized with stochastic pa-
rameter ϑ can be expressed as a weighted sum of the ZCBs evaluated at a given realization of
ϑ and where ω1 + · · · + ωN = 1. Note that under the HJM framework, each ZCB, P (t, T ; θn )
in (2.9) is arbitrage-free; therefore, the convex combination is too. This can be shown by
checking whether P (S, T ; {θn }N n=1 )/M (S) is indeed a martingale for some t < S < T :
"P # N
N
P (S, T ; {θn }N
   
n=1 ) n=1 ω n P (S, T ; θ n ) X P (S, T ; θn )
Et = Et = ωn Et ,
M (S) M (S) n=1
M (S)
5
since every discounted ZCB, P (S, T ; θn )/M (S), is a martingale under the Q measure we have,
N
P (S, T ; {θn }N P (t, T ; {θn }N
  X
n=1 ) P (t, T ; θn ) n=1 )
Et = ωn = .
M (S) n=1
M (t) M (t)

This result shows that the application of the RAnD method enables the transition from a
continuous random variable to a discrete one; the cost of this transition is expressed by N .
Moreover, as shown above, the error N in (2.9) does not impact the arbitrage conditions-
the pricing under the RAnD method is simply a weighted average of arbitrage-free prices.
However, from a practical perspective, keeping this error as small as possible is still essential.
This is particularly important when calibrating the randomized model to market data, i.e., the
calibration procedure is associated with varying parameters of the variable ϑ. It is, therefore,
important that the discretized version resembles the continuous version as closely as possible.
In the numerical section, Section 4.5, we will investigate the convergence aspects of different
randomizers and an optimal number of expansion terms, N .

2.2. The randomized Hull-White (rHW) short-rate model

This section provides a specification leading us to the randomized version of the famous
HW model [14]-a single-factor, no-arbitrage yield curve model, in which an extended Ornstein-
Uhlenbeck drives the short-term interest rate mean-reverting process. Under the HJM frame-
work and the arbitrage-free condition for the drift in (2.1), the rHW model is specified by:

γ(t, T ) = η · e−λ(T −t) , t < T, (2.10)

where we consider three different randomization cases: the randomization of the volatility
parameter, η, the mean-reversion, λ, or the randomization of both parameters using bivariate
distribution:
d d d
η = ϑ1 , or λ = ϑ2 , or λ|η = ϑ2 |ϑ1 . (2.11)
To simplify the notation, we consider one parameter θ ∈ {η, λ} with the corresponding random
variable ϑ. Particular choices of the parameters will be given explicitly.
Given the HJM volatility in (2.10), under the risk-free measure, the dynamics of the HW
model read:
dr(t) = λ(ψ(t) − r(t))dt + ηdW (t), r0 ≡ f (0, 0), (2.12)
with
1 η2 ∂ log P (0, t)
f (0, t) + 2 1 − e−2λt , f (0, t) = −


ψ(t) = f (0, t) + , (2.13)
λ 2λ ∂t
where ψ(t) is a time-dependent drift term, which ensures the fit of the calibration to a yield
curve observed in the market, W (t) is the Brownian motion under measure Q, and f (0, t)
indicates the instantaneous forward rate computed from the yield curve, defined in terms
of ZCBs, P (0, t). Parameter η determines the overall level of the volatility, and λ is the
reversion rate parameter. A large value of λ causes short-term rate movements to dampen out
rapidly, reducing the long-term volatility. Because of this interdependence between the model
parameters, often, in practical applications, λ is fixed and set to a constant. In contrast, η is
often set to be piece-wise constant and calibrated such that the ATM volatilities are calibrated.
We will show that under the RAnD method, such a strategy is inadequate.
2
The short-rate r(t) in (2.12) is thus normally distributed with r(t) ∼ N (µr(t) , σr(t) ), with

t
η2
Z
µr(t) := r0 e−λt + λ ψ(z)e−λ(t−z) dz, 2
1 − e−2λt .


σr(t) := (2.14)
0 2λ

Moreover, for ψ(t) constant, i.e., ψ(t) ≡ ψ (in this case we deal with the Vašiček model [18]),
we have limt→∞ Et0 [r(t)] = ψ. This means that the first moment of the process converges to
the mean-reverting level ψ, for large values of t.
As a first step, we check the impact of randomization on the probability density function,
PDF, of the interest rate process r(t) in (2.12). Since the HW model belongs to the affine class

6
of processes, we can benefit from the available ChF. Given a stochastic parameter ϑ, the ChF
is given by:
h RT i h h RT ii
φr (u; t, T ) := Et e− t r(s)ds+iur(T ) = Et Et e− t r(s)ds+iur(T ) ϑ = θ .

By application of the quadrature rule in Theorem 2.1 and Appendix A.2, we find:
Z N
X
φr (u; t, T ) = φr|ϑ=θ (u; t, T )dFϑ (θ) = ωn φr(T )|ϑ=θn (u; t, T ) + F
N, (2.15)
[a,b] n=1

and by utilizing the Fourier transform, the PDF of the rHW model reads,
Z N N
1 X X
fr(T ) (x) = e−iux ωn φr|ϑ=θn (u; t, T )du + F
N = ωn fr(T ;θn ) (x) + F
N,
2π R n=1 n=1

where r(t)|ϑ = θn indicates the HW model with a particular realization, θn , of the randomized
variable. The representation above presents the relation between the densities of the rHW
process, r(t), as a convex combination of the HW processes, r(t)|θn with a particular parameter
θn . We also highlight that the error F N in (2.15) may differ from the quadrature error in (2.6).
Because of different randomization choices, we distinguish the following randomization
types and their associated error:
d

XN 
 fr(T ) (y) + N , for η = ϑ,
d
fr(T ) (y) = ωn fr(T ;θn ) (y) + F
N =: fre(T ) (y) + e
N , for λ = ϑ, (2.16)
d d

n=1 
frb(T ) (y) + bN , for η = ϑ1 & λ|η = ϑ2 .
Under the RAnD method, the PDF, fr(T ) (y) in (2.16) is given as a linear combination of normal
densities that depend on different parameter realizations θn , resembling a similar problem as
presented in [3].
In the next section, we associate the PDF given in (2.16) and find the corresponding SDE
for a 1D stochastic representation.

2.3. Dynamics of the rHW model

The combination of Equations (2.11) and (2.12) shows us that the simulation of the ran-
domized short-rate models is explicit using the Monte Carlo technique. However, by an explicit
form of the corresponding SDE, we will get more insight into the randomization and its impact
on the dynamics of the driving process.
As a start, under the rHW model, consider the randomization of the volatility parameter,
θ := η, with ϑ(ω) = θ, and where the corresponding quadrature pairs are given by {ωn , ηn }N
n=1 .
For such a setting, under each realization ηn , we have the associated HW model, rn (t) :=
rn (t; ηn ), with the dynamics given by:
drn (t) = λ(ψ n (t) − rn (t))dt + ηn dW (t), rn (t0 ) = f (0, 0), n = 1, . . . , N, (2.17)
with common Brownian motion, W (t), and initial value rn (t0 ) = f (0, 0) in (2.13), where λ
is constant and equal for all the rn (t), and where ψ n (t) is defined in (2.13) and is given as a
function of λ and ηn .
Given the sequence of HW model processes, r1 (t), . . . , rN (t), in (2.17) and the probability
density relation in (2.16), we consider the problem of finding the corresponding SDE for the
rHW process, r(t). Formally, we seek an SDE, with the solution given in (2.16) and where
each of the constituent processes, rn (t), is driven by (2.17). Thus, we consider the following
process,
dr(t) = λ(t, r(t))dt + η(t, r(t))dW (t), r(t0 ) = f (0, 0), (2.18)
with some state-dependent drift, λ(t, r(t)), and volatility, η(t, r(t)), and where Brownian mo-
tion W (t) is as in (2.17). As indicated in [2], the problem of finding an SDE when marginal
distributions and the corresponding weights are given is the reverse of finding the marginal
density function of the solution of an SDE when the coefficients are known. Moreover, the
problem of a mixture of normal or lognormal processes is known [2]. Here, in Proposition 2.1,
we adapt those techniques and provide an explicit form for the SDE in (2.18).
7
Proposition 2.1 (Local volatility process for the HW model with randomized volatility pa-
rameter, η). Let us assume a sequence of positive constants ηn , n = 1, . . . , N. Then, the SDE

dr(t) = λ(t, r(t))dt + η(t, r(t))dW (t), r(t0 ) = f (0, 0), (2.19)
with
N
X N
X
λ(t, y) = Λn (t, y)λ(ψ n (t) − y), η 2 (t, y) = ηn2 Λn (t, y), (2.20)
n=1 n=1

where:
ωn fr(t;ηn ) (y)
Λn (t, y) = PN ,
n=1 ωn fr(t;ηn ) (y)

has a strong solution whose marginal density is given by the mixture of normal probability
density functions:
N
X
fr(t) (y) = ωn fr(t;ηn ) (y), (2.21)
n=1

PN
where n=1 ωn = 1 for ωn ≥ 0, n = 1, . . . , N with fr(t;ηn ) (x) the PDF of the HW model with
dynamics, given by:

drn (t) = λ(ψ n (t) − rn (t))dt + ηn dW (t), rn (t0 ) = f (0, 0), (2.22)
2
ηn
where rn (t) := rn (t; ηn ) with ψ n (t) = f (0, t) + λ1 f (0, t) + 1 − e−2λt .


2λ2

Proof. The problem we address is the derivation of the drift, λ(t, y), and the volatility function,
η(t, r(t)), in the SDE:
dr(t) = λ(t, r(t))dt + η(t, r(t))dW (t), (2.23)
such that
N
X
fr(t) (y) = ωn fr(t;ηn ) (y), (2.24)
n=1

where fr(t;η) (y) is the PDF of the Hull-White process in (2.17), with parameter η. Under the
HW model dynamics, the so-called linear-growth condition holds, i.e., ηn2 ≤ Cn (1 + y 2 ) for
i = 1, . . . , N. Assuming that the same non-explosion condition holds for (2.23), i.e., η 2 (t, y) ≤
C(1 + y 2 ), uniformly in t, we start by deriving the Fokker-Planck equation for r(t):

∂ 1 ∂2 ∂
η 2 (t, y))fr(t) (y) −


fr(t) (y) = 2
λ(t, y)fr(t) (y), (2.25)
∂t 2 ∂y ∂y

while for each individual interest rate process, rn (t), in (2.22) we have:

∂ 1 ∂2 ∂
ηn2 fr(t;ηn ) (y) −



fr(t;ηn ) (y) = 2
λ(ψ n (t) − y) fr(t;ηn ) (y). (2.26)
∂t 2 ∂y ∂y

After substituting (2.24) into (2.25):

N
" N
# N
∂ X 1 ∂2 2
X ∂ X
ωn fr(t;ηn ) (y) = η (t, y) ωn f r(t;η ) (y) − λ(t, y) ωn fr(t;ηn ) (y).
∂t n=1 2 ∂y 2 n=1
n
∂y n=1

Due to the linearity of the derivative operator, we find,

N
" N
# N
X ∂ 1 ∂2 2
X ∂ X
ωn fr(t;ηn ) (y) = η (t, y) ωn f r(t;η ) (y) − λ(t, y) ωn fr(t;ηn ) (y).
n=1
∂t 2 ∂y 2 n=1
n
∂y n=1

8
By substitution of (2.26) and operator linearity,
N N
∂2 X 1 2 X ∂
ωn η fr(t;ηn ) (y) − ωn λ(ψ n (t) − y)fr(t;ηn ) (y) =
∂y 2 n=1 2 n n=1
∂y
" N
# N
1 ∂2 2
X ∂ X
η (t, y) ω f
n r(t;ηn ) (y) − λ(t, y) ωn fr(t;ηn ) (y).
2 ∂y 2 n=1
∂y n=1

Finally, by matching the appropriate terms and integration, we find:

N
X N
X
ωn ηn2 fr(t;ηn ) (y) = η 2 (t, y) ωn fr(t;ηn ) (y) + C1 (t)y + C2 (t),
n=1 n=1
N
X N
X
ωn λ(ψ n (t) − y)fr(t;ηn ) (y) = λ(t, y) ωn fr(t;ηn ) (y) + C3 (t),
n=1 n=1

for some time-dependent functions C1 (t), C2 (t) and C3 (t). Since the LHS and RHS of the
equations above need to satisfy uniform convergence requirements implying that for y → ∞,
they converge to 0, the following needs to hold: C1 (t) = C2 (t) = C3 (t) = 0, ∀t. Therefore, the
expressions for λ(t, y) and η(t, y) read:
PN 2
PN
n=1 ωn ηn fr(t;ηn ) (y) ωn λ(ψ n (t) − y)fr(t;ηn ) (y)
2
η (t, y) = PN , λ(t, y) = n=1 PN . (2.27)
n=1 ωn fr(t;ηn ) (y) n=1 ωn fr(t;ηn ) (y)

By setting
ωn fr(t;ηn ) (y)
Λn (t, y) = PN for n = 1, . . . , N, (2.28)
n=1 ωn fr(t;ηn ) (y)
we can write
N
X N
X
λ(t, y) = Λn (t, y)λ(ψ n (t) − y), and η 2 (t, y) = ηn2 Λn (t, y). (2.29)
n=1 n=1
PN
Finally, by taking η∗ := maxi=1,...,N ηn and since n=1 Λn (t, y) = 1, ∀y, we have:
N
X N
X
η 2 (t, y) = ηn2 Λn (t, y) ≤ η∗2 Λn (t, y) = η∗2 = C.
n=1 n=1

Since the volatility parameter η 2 (t, y) is bounded by a constant, the uniform convergence
criterion is satisfied. The uniqueness of the strong solution follows from Theorem 12.1 in [17]
while in [2] (Theorem 2.1), a proof for a generic case for a normal mixture is provided. 
Proposition 2.1 illustrates that under the randomized volatility parameter for the HW
model, the normal mixture dynamics resemble a one-dimensional local-volatility-type diffusion
process. The local volatility function is expressed as a weighted volatility squared, ηn2 , of the
constituent processes, rn (t), and where the weights are functions of the quadrature coefficients,
ωn , and the corresponding PDFs, fr(t;ηn ) (y).
Following the same strategy, we derive the dynamics of a process of the rHW model with
the randomized mean-reversion parameter, λ. Proposition 2.2 provides the details.
Proposition 2.2 (Dynamics of the HW model with randomized mean-reversion parameter,
λ). Let us assume a sequence of positive constants λm , m = 1, . . . , M , then the SDE

de
r(t) = λ(t,
e re(t))dt + ηe(t, re(t))dW (t), re(t0 ) = f (0, 0), (2.30)

with ηe(t, y) = η, and

M
e m (t, y) = P $m fre(t;λm ) (y)
X
λ(t,
e y) = e m (t, y)λm (ψm (t) − y), where Λ
Λ , (2.31)
M
m=1 m=1 $m fr e(t;λm ) (y)

9
has a strong solution whose marginal density is given by the mixture of normal probability
density functions:
M
X
fre(t) (x) = $m fre(t;λm ) (x), (2.32)
m=1
PM
where m=1 $m = 1 for $m ≥ 0, m = 1, . . . , M , with fre(t;λm ) (x) is the PDF the HW model
whose dynamics is given by:

rm (t) = λm (ψm (t) − rem (t))dt + ηdW (t), rem (t0 ) = f (0, 0),
de
η2
1
1 − e−2λm t .


where ren (t) := ren (t; λn ) with ψm (t) = f (0, t) + λm f (0, t) + 2λ2m

Proof. The proof is analogous to Proposition 2.1. 

With the rHW model, the randomization of either model parameters enables us to determine
the corresponding 1D SDE. For the random volatility parameter, η, we derived a local-volatility
type short-rate process; however, in the case of random λ, the corresponding 1D process has
a different structure, i.e., the volatility coefficient, ηe(t, y) = η, in (2.30), stays constant. When
analyzing the dynamics of the corresponding ZCB (see Corollary 2.1), which is a tradable asset,
the dynamics will resemble a local-volatility type process. To derive the corresponding SDE,
one needs to follow the same strategy as in Propositions 2.1 and 2.2 with the process for the
ZCB, instead of the short-rate process.
In the final part of this section, the case of joint randomization of both HW model param-
eters will be discussed.

2.3.1. Dynamics of the rHW model with bivariate distribution for λ and η.
An extension of the RAnD method is to consider both HW model parameters random
and follow a bivariate distribution. Such an extension may benefit from the interconnection
between model parameters and possibly the correlation coefficient; however, it would require
that the conditional moments are known explicitly. This is troublesome because only for a few
random distributions the moment functions are known in the closed form. However, if we stay,
for example, within the Gaussian world, such an extension to the 2D case is possible.
Corollary 2.2 (Random parameters with bivariate distribution). Under a bivariate distribu-
tion Θ = [ϑ1 , ϑ2 ] with ζ(ϑ1 ) = {ω1,n , θ1,n }N M
n=1 and conditioned on ζ(ϑ2 |ϑ1 ) = {ω2,m , θ2,m }m=1 ,
the randomized ChF is given by:
N
X M
X
φX (u; t, T ) = ωn ωm φX|ϑ1 =θn ,ϑ2 =θm (u; t, T ) + bN,M , (2.33)
n=1 m=1

where N and M indicate the number of expansion terms for ϑ1 and ϑ2 |ϑ1 respectively, ϑ2 |ϑ1
indicates a conditional random variable, bN,M is the corresponding aggregated error, and the
remaining specification follows Theorem 2.1.
The PDF computation for the bivariate case requires the sequential computation of the
associated weights and the corresponding points, which can be established utilizing Proposi-
tions 2.1 and 2.2. In the first iteration step, we compute the grid points associated with the
volatility parameter, η, and for each realization ηn , we establish the corresponding conditional
PDF. By summing over all possible pairs, the unconditional PDF for the rHW model can be
computed,
M
X N X
X M
fr(T ;ηn ) (x) = $m,n fr(T ;ηn ,λm ) (x), fr(T ) (x) = ωn $m,n fr(T ;ηn ,λm ) (x), (2.34)
m=1 n=1 m=1

where $m,n is defined in (2.32), ωn is given in (2.21) and fr(T ;ηn ,λm ) (x) indicates the PDF of
the HW model with parameters ηn and λm , and is defined in (2.14). Proposition 2.3 provides
the dynamics of the associated rHW model.
10
Proposition 2.3 (Local volatility process for the HW model with randomized parameters).
Let us assume a sequence of positive constants ηn , n = 1, . . . , N , and λm,n , m = 1, . . . , M ,
then the SDE:
db
r(t) = λ(t,
b rb(t))dt + ηb(t, rb(t))dW (t), rb(t0 ) = f (0, 0), (2.35)
with
N X
X M N X
X M
ηb2 (t, rb(t)) = b n,m (t, y)ηn2 , λ(t,
Λ b rb(t)) = b n,m (t, y)λm,n (ψm,n (t) − y),
Λ
n=1 m=1 n=1 m=1

where
ωn $m,n fr(t;ηn ,λm ) (y)
Λ
b n,m (t, y) = P
N PM ,
n=1 m=1 ωn $m,n fr(t;ηn ,λm,n ) (y)
has a strong solution whose marginal density is given by the following mixture of normal prob-
ability density functions:
N X
X M
frb(t) (y) = ωn $m,n fr(t;ηn ,λm,n ) (y),
n=1 m=1
PN PM
where n=1 ωn = 1, m=1 $m,n = 1 for ωn , $m,n ≥ 0, n = 1, . . . , N , m = 1, . . . , M with
fr(t;ηn ,λm,n ) (y) being the PDF of the HW model whose dynamics are given by:

drm,n (t) = λm,n (ψm,n (t) − rm,n (t))dt + ηn dW (t), rm,n (t0 ) = f (0, 0),
2
ηn
1
1 − e−2λm,n t .


and rn,m (t) := r(t; ηn , λm ) with ψm,n (t) = f (0, t) + λm,n f (0, t) + 2λ2m,n

Proof. The proof is analogous to Proposition 2.1. 

The concept of multivariate model parameters can be extended further. In the case of
the HW model with piece-wise constant parameters, it is possible to randomize each piece-
wise element, to mimic the parameter controlled by the stochastic process. Unfortunately, as
the number of ranges increases, the number of terms in the summation in (2.34) will grow
exponentially.
The impact of bivariate model parameters on implied volatilities will be analyzed further
in Section 4.

3. Pricing under the randomized Hull-White (rHW) model

This section focuses on the pricing under the rHW model. The presented pricing equations
utilize Theorem 2.1. Throughout the section, we denote by VHW (t, r(t; θ)) with θ ∈ {η, λ}, the
Hull-White price of a derivative depending on the short-rate r(t; θ) with parameters λ and η,
by VrHW (t, r(t; ϑ)) we denote the pricing under the rHW model.

3.1. Swaptions under rHW model

While in Equation (2.5), there is a relationship between HW and rHW prices, here explicit
pricing formulas will be given. The valuation of swaption contracts in the rHW model can be
carried out at the same complexity level as in the standard HW model.
The critical ingredient for pricing swaptions under the rHW model is an analytical expres-
sion for the options of the pure ZCB. Lemma 3.1 provides the pricing equations under the rHW
model.
Lemma 3.1 (Pricing of options on a ZCB under the rHW model). Under the randomized HW
model parameters, θ := η or θ := λ, with ϑ(ω) = θ, the price of a European-style option on a
ZCB, P (T, S), with option’s expiry, T , strike, K, and maturity of the underlying bond S, is
given by:
N
X
VχZ (t, T, S, K; ϑ) = ωn VχZ (t, T, S, K; θn ), (3.1)
n=1

11
where,
 
M (t)
VχZ (t, T, S, K; θn ) = EQ
t max(χ(P (T, S; θn ) − K), 0)
M (T )
= χP (t, S)FN (0,1) (χdn ) − χKP (t, T )FN (0,1) (χ(dn − σ̄n )), (3.2)

with P (t, S) and P (t, T ) the ZCB computed from the associated yield curve; P (T, S; θn ) is the
ZCB obtained from the model with the randomized parameter, ϑ; χ = 1 and χ = −1 corresponds
to call and put options, respectively. Depending on the randomized parameter, we have:

ηn2  
for η : σ̄n2 = 1 − e−2λ(T −t) B 2 (T, S; λ), (3.3)
2λ
η2  
for λ : σ̄n2 = 1 − e−2λn (T −t) B 2 (T, S; λn ), (3.4)
2λn
with
1 P (t, S) σ̄n 1 
dn = log + , B(T, S; λ) = 1 − e−λ(S−T ) ,
σ̄n P (t, T )K 2 λ
where the HW model is defined in (2.12) and FN (0,1) (·) corresponds the standard normal CDF.
Proof. The proof is a direct consequence of combining Theorem 2.1 with the pricing of options
on a ZCB under the HW model (as given in [4, 16]). 
Now, we consider an interest rate swap with a fixed rate, K, and payment times T =
{Ti , Ti+1 , . . . , Tm } and the corresponding reset rates {Ti−1 , Ti+1 , . . . , Tm−1 } with the payoff
given by:
m
X
Swap
HP/R (T , K) = ᾱ τk (`(Tk−1 ; Tk−1 , Tk ) − K) ,
k=i

for τk = Tk − Tk−1 , with P indicating a swap payer for ᾱ = 1 and a swap receiver for ᾱ = −1
and where `(t; Tk−1 , Tk ) stands for the libor rate over the period [Tk−1 , Tk ] observed at time t.
To determine today’s value of the swap, we evaluate the corresponding expectation of the
discounted future cash flows, i.e., each payment which takes place at the time points, Ti , . . . , Tm ,
needs to be discounted to today,
m   m
Swap
X M (t)
X

VP/R (t, K, T ) = ᾱ τk EQ
t `k (Tk−1 ) − K = ᾱ τk P (t, Tk ) `k (t) − K , (3.5)
M (Tk )
k=i k=i

with `k (t) := `(t; Tk−1 , Tk ).

We derive the valuation formula for a swaption contract with strike K and option expiry
T = Ti−1 . Using (3.5), the pricing equation reads:
 
Q M (t)
 h  i
Swpt Swap Swap
VP/R (t, T, K, T ) = Et max VP/R (T, K, T ), 0 = P (t, T )ETt max VP/R (T, K, T ), 0 ,
M (T )

where the expectation, ETt [·], is taken under the T −forward measure and where the price of a
swap price at T = Ti−1 is given by:
" m
# m
Swap
X X
VP/R (T, K, T ) = ᾱ 1 − P (T, Tm ) − K τk P (Ti , Tk ) = ᾱ − ᾱ ck P (Ti , Tk ),
k=i k=i

with ck = Kτk for k = i, . . . , m − 1, cm = 1 + Kτm , and τk = Tk − Tk−1 .

So, the pricing equation becomes:
" m
!#
Swpt
X
T
VP/R (t, T, K, T ) = ᾱP (t, T )Et max 1 − ck P (Ti , Tk ), 0 .
k=i

12
 Up to this point, the pricing equations do not depend on specific model choices. Now,
however, we consider pricing under the rHW model, which, via the conditional expectation
approach, as presented in Lemma 2.1, yields:
" " m
##
 X 
Swpt T T
VP/R (t, T, K, T ; ϑ) = ᾱP (t, T )Et Et max 1 − ck P (Ti , Tk ), 0 ϑ = θ (3.6)
k=i
N
" m
#
X  X 
= ᾱP (t, T ) ωn ETt max 1 − ck P (Ti , Tk ; θn ), 0 + N .
n=1 k=i

Here, N is the associated quadrature error defined in (2.6). Since the inner expectation
in (3.6) resembles the expression for a swaption under the classical HW, we follow the standard
procedure and apply the so-called Jamshidian trick [15] to exchange the maximum operator
and the expectation. The resulting pricing equations for swaptions under the rHW model are
given in Lemma 3.2.
Lemma 3.2 (Pricing of Swaptions under randomized Hull-White model). Consider the rHW
model, with parameters {λ, η} and the randomizing random variable ϑ, which randomizes either
of the model parameters. For a unit notional, a constant strike, K, option expiry T = Ti−1
and a strip of swap payments T = {Ti , . . . , Tm }, with Ti > Ti−1 and accruals τi = Ti − Ti−1 ,
the prices of swaption payer and receiver, P/R := Payer/Receiver, are given by:
N
X m
X
Swpt
VP/R (t0 , T, T , K; ϑ) = ωn ck VχZ (t0 , T, Tk , K̂k (θn ); θn ), (3.7)
n=1 k=i

with a swaption payer, P, for χ = −1, swaption receiver, R, with χ = 1, where VχZ (·) is defined
in (3.2) and where the strike price K̂k (θn ) = exp (A(T, Tk ; θn ) + B(T, Tk ; θn )rn∗ ). Here, rn∗ is
determined by solving, for each parameter realization θn , the following equation:
m
X  
1− ck exp A(T, Tk ; θn ) − B(T, Tk ; θn )rn∗ = 0, n = 1, . . . , N, (3.8)
k=i

where
P (0, Tk ) η2
1 − e−2λT B 2 (T, Tk ; {λ, η}) ,


A(T, Tk ; {λ, η}) = log + B(T, Tk ; {λ, η})f (0, T ) −
P (0, T ) 4λ
1 
B(T, Tk ; {λ, η}) = 1 − e−λ(Tk −T ) ,
λ
with ck = Kτk for k = i, . . . , m − 1, cm = 1 + Kτm and where the pairs {ωn , θn }, n = 1, . . . , N ,
are based on the randomizer ϑ and computed using Appendix A.2.
Under the rHW model, the pricing Equation (3.7) for swaptions is a direct application
of Theorem 2.1. The pricing under randomized parameters is simply an average of non-
randomized prices, with varying model parameters, accompanied by weights based on the
randomizer.
From the computational perspective, the swaption pricing in (3.7) requires N swaption
prices for different realizations θn , n = 1, . . . , N ; therefore, N optimization problems in (3.8)
need to be solved. This computational complexity can be greatly reduced by employing the
multi-d Newton-Raphson algorithm to determine optimal rn∗ for n = 1, . . . , N.
Remark 3.3 (Computation of sensitivities). The computation of sensitivities under the RAnD
method is straightforward, i.e., because the pricing is expressed as a convex combination, the
sensitivities are expressed as a weighted sum of individual derivatives. For example, in the case
of the swaption pricing, the sensitivity to a particular market quote q is expressed as a sum of
sensitivities of individual options on ZCBs:
N M
∂ Swpt X X ∂
VP/R (t0 , T, T , K; ϑ) = ωn ck VχZ (t0 , T, Tk , K̂k (θn ); θn ), (3.9)
∂q n=1
∂q
k=i

13
with the specification as given in Lemma 3.2.
On the other hand, the sensitivity to the quadrature pairs, {ωn , θn }, ∂θn /∂â and ∂ωn /∂â
may be, for some specific cases, computed analytically (see Section 4.3). Still, in a generic
setting, it is recommended to compute these derivatives numerically, with, for example, finite
differences:  
ζ(ϑ(â + δâ )) − ζ(ϑ(â − δâ )) ∂ωn ∂θn
≈ , ,
2δâ ∂â ∂â
where ϑ(â) indicates the dependence of the random variable and parameter â, δâ is the shock
size and ζ(ϑ) : R → {ωn , θn }N n=1 is defined in Appendix A.2. Due to the applied finite
difference shocks to â, an additional bias will be introduced. We expect, however, this error to
be of acceptable magnitude, as is commonly observed in current financial practice.

4. Numerical experiments

In this section, several pricing experiments will be performed. First, we present a detailed
analysis of the rHW model in realistic pricing scenarios. In the first experiment, we explore
the implied volatility smile evolution in time, comparing the implied volatility surface from the
HW and rHW models. As the next step, the study of the parameter randomization on shapes
of implied volatilities will be illustrated. Finally, the calibration results with market data will
be presented in the conclusive experiment. This section will end with numerical experiments
involving bivariate distribution for the model parameters.

4.1. Evolution of implied volatilities

We analyze the swaption implied volatilities observed in the market and compare them to
implied volatilities generated using the HW and rHW models. We focus here on the flexibility
in generating realistic implied volatility shapes. Throughout the section, we will consider the
shifted implied volatilities computed by inverting Black’s formula, which, for ᾱ ∈ {1, −1} for
call and put options, respectively, reads:

VB (T, K, F0 , σs , ᾱ, s) = ᾱ · (F0 + s) · FN (0,1) (ᾱd1 ) − ᾱ(K + s)FN (0,1) (ᾱd2 ), (4.1)
1 
√ log(F0 + s)/(K + s) + 1/2σs2 T ,

d1 =
σ T
√
d2 = d1 − σs T ,

with shift parameter s, F0 being the forward rate, σs is the corresponding volatility coefficient,
K being the strike, and T corresponds to the time to option expiry.
Then, the shifted Black’s formula for swaptions is given by:
m
B,Swpt
X P (t0 , Ti−1 ) − P (t0 , Tm )
VP/R = VB (T, K, S(t0 ), σs , ᾱ, s) τi P (t0 , Ti ), S(t0 ) = Pm , (4.2)
i=1 i=1 τi P (t0 , Ti )

with swaption payer, P , for ᾱ = 1 and swaption receiver, R, for ᾱ = −1 and where S(t0 ) is
the corresponding swap rate. The shift parameter, s, is typically perceived as the lower bound
for the interest rates by the market participants and varies depending on the currency. It is
important to note that when inverting the Black’s formula in (4.2), the corresponding implied
volatility σs is a function of the shift parameter, i.e., for different choices of s, the implied
volatilities are different. In principle, this is not a problem as long as the implied volatilities
from the market and model volatilities are computed with the same shift coefficient.
In the first experiment, we compare the implied volatility surfaces of the HW model and
the rHW model. In the rHW model, the speed of mean-reversion is randomized by a uniform
distribution, λ ∼ U([â, b̂]), on an interval [â, b̂]. The numerical results are illustrated in Figure 1.
The results demonstrate that the HW model can only generate implied volatility skew. At the
same time, the randomization of the mean-reversion parameter, λ, shows implied volatility
skew and smile. Although the randomization is not time-dependent, i.e., the parameters are
stochastic but stationary, we observe a time evolution of the implied volatilities. The same
phenomenon has been observed in [3, 11], where the randomized Black-Scholes model was
14
 Figure 1: Swaption volatility evolution for the HW and rHW models implied by the
shifted Black’s model. The simulation was performed for varying swaption option expiry,
T , and a fixed tenor of 1y. The parameters specified in the experiment are: for the HW
model: η = 0.005, λ = 0.001 and for the rHW model: η = 0.005 and λ ∼ U ([−0.15, 0.6]).
In the experiment, the implied volatilities are computed with zero shift parameter, s = 0.

considered. The reported numerical results are promising. By a stochastic mean-reversion

parameter, the HW model has an additional degree of freedom that improves the model’s
flexibility. In the following sections, we will analyze the randomization effect of all model
parameters and the different choices of randomizing distributions.

4.2. Randomization and impact on implied volatilities

Here, we study the impact of different randomization choices on the implied volatility shape.
Understanding how parameters affect the generated volatilities is crucial in model calibration.
A clear relationship between the model parameters and the associated skew, curvature and
volatility level is fundamental, enabling traders to react swiftly to market movements. It is
also the basis for setting up a calibration routine when a particular market movement can be
encapsulated in a parameter change.
In the first experiment, we analyze the randomization of volatility parameter η. We consider
two randomization cases where η either follows a uniform or a normal distribution. The details
regarding the computation of the associated moments and the corresponding quadrature points
can be found in Appendix A. The experiment is set up so that both approaches’ mean values,
E[η], are equal. The results are presented in Figure 2, where it is shown that the randomization

Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
70 70

60 60

50 50
Implied Volatility, [%]

Implied Volatility, [%]

40 40

30 30

20 20

10 10

0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)

Figure 2: Impact of randomized volatility parameter η on implied volatilities with a fixed

speed of mean-reversion parameter λ = 0.009. Left: η ∼ U ([â, b̂]). Right: η ∼ N (â, b̂).

of the volatility parameter η has a pronounced effect on the level of implied volatility with
minimal effect on the smile. The curvature is visible for uniform randomization. It is essential

15
to note that a higher variance for the normal distribution may give rise to a higher curvature
but may also cause issues related to negative volatilities. Our experiments have shown that
even for distributions defined in the positive domain, the impact on the smile is limited, even
for fat-tailed random variables.
A much richer spectrum of implied volatility shapes is obtained when the randomization
technique is applied to the mean-reversion parameter λ. Figure 3 presents the randomization
with either uniform or normal random variables. A substantial amount of curvature can be
generated by taking the mean-reversion random. We also report that the curvature change
affects the overall volatility level, i.e., it is impossible to keep the level fixed and only adjust the
smile. However, the implied volatility level can be fixed by adjusting the volatility parameter
η. This strategy will be discussed further in the context of model calibration.

Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
20 18

18 16

16 14
Implied Volatility, [%]

Implied Volatility, [%]

14 12

12 10

10 8

8 6

6 4
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)

Figure 3: Impact of randomized volatility parameter λ on implied volatilities with a

fixed volatility parameter η = 0.0050. Left: λ ∼ U([â, b̂]). Right: λ ∼ N (â, b̂).

In the final experiment of this section, we consider the randomization of λ ∼ U([â, b̂]) using
uniform distribution and check how the parameters â and b̂ affect the implied volatilities. In
Figure 4, the results show an interesting pattern: the curvature level is mainly driven by the
distance |b̂ − â|, i.e., the larger the distance, the more implied volatility smile is generated.
Changes of either of the parameters affects the volatility level; therefore, some of the volatility
effect, η, can be offset by the interval [â, b̂].

Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
18 18

16
16

14
14
Implied Volatility, [%]

Implied Volatility, [%]

12

12

10

10
8

8
6

6 4
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)

Figure 4: Impact of randomized volatility parameter λ ∼ U ([â, b̂]) on implied volatilities

with a fixed volatility parameter η = 0.0050. Left: varying parameter â. Right: varying
parameter b̂.

Given the numerical results presented above, we conclude that the most considerable impact
on implied volatilities comes from the mean-reversion, λ, and not from the volatility parameter,
η. This is particularly interesting when we confront these results with Proposition 2.1 and
Proposition 2.2, where the dynamics of the corresponding short-rate processes were derived
16
and where it was shown that randomization of the mean-reversion did not lead to the local
volatility type of dynamics. In contrast, the randomization of η does, which can be explained
by the nature of interest rate derivatives, where derivative prices are driven by the dynamics
of the ZCBs, but not directly by the short-rate process. However, since under the HW model,
the volatility of the ZCBs is given by both parameters (see [16]), the randomization of either
of them will imply a local volatility type of dynamics for the ZCBs.

4.3. Calibration of the randomized Hull-White model

Every model that may be considered successful must also show its ability to calibrate to
the market quotes. Moreover, the calibration and pricing process needs to be computationally
efficient, especially when considering the pricing of large portfolios, for example, in the context
of xVA. Therefore, this section focuses on calibrating the rHW model to swaption implied
volatilities.
Calibration of model parameters requires multiple iterations over the parameter space until
specific optimization criteria are met. This implies that the pricing needs to be repeated at
every choice of parameter candidate. By randomization of the model parameters, the number
of free parameters will increase. The additional degrees of freedom will depend on the type of
randomizer chosen. Let us consider, for example, parameter randomization with ϑ(â, b̂), which
is driven by two model parameters â and b̂. Two parameters of the randomizer ϑ correspond
to one additional degree of freedom that can be used for model calibration.
From the computational perspective, using the RAnD method requires that for every iter-
ation step in the calibration procedure, the mapping between the randomizing variable ϑ and
the corresponding pairs {ωn , θn }, n = 1, . . . , N , needs to be employed. Although the algorithm
given in Appendix A.2 is straightforward (it mainly depends on the computation of eigenval-
ues), it can be further simplified. When randomizing variables can be expressed as a linear
combination of some base random variable, for example, normal or uniform, the calculations
of the weights and the corresponding nodes can be significantly simplified. In the case of a
normal randomizer, ϑ ∼ N (â, b̂2 ), we can benefit from the linearity of the normal distribution.
For the standard normal and its associated points, we have θn = â+ b̂·θN ,n , where θN ,n are the
nodes corresponding to standard normal, N (0, 1). This implies that we can simply tabulate the
results for the standard normal and scale the points accordingly. In the case of the weights ωn ,
they stay invariant to a linear transformation [12]. A similar property holds for ϑ ∼ U([â, b̂]),
where the nodes can be computed for ϑ ∼ U([0, 1]) and scaled appropriately.
In the calibration experiment, we consider the market data for the USD market as of the
18th of August, 2022. In all calibration exercises, we consider a fixed tenor of 1y and analyze
the accuracy for varying expires and strikes. The calibrated model parameters for both HW
and rHW models are tabulated in Table 1. The results are intriguing, i.e., in the randomization
for λ, we were able to calibrate all swaptions while keeping the mean of the randomizer fixed at
0.1. This shows that having more degrees of freedom is not necessary to improve the calibration
results. As in the standard HW model, we have only used two parameters.

Table 1: Calibration of the HW and rHW model: parameters determined in swaption calibra-
tion.

Hull-White RAnD Hull-White

T , expiry η λ η λ
1y 0.0094 0.0090 0.0091 λ ∼ N (0.1, 0.452 )
2y 0.0082 0.0035 0.0080 λ ∼ N (0.1, 0.332 )
5y 0.0069 0.0020 0.0079 λ ∼ N (0.1, 0.162 )
8y 0.0067 0.0095 0.0080 λ ∼ N (0.1, 0.122 )
10y 0.0067 0.0090 0.0082 λ ∼ N (0.1, 0.112 )
15y 0.0064 0.0080 0.0085 λ ∼ N (0.1, 0.092 )
20y 0.0060 0.0080 0.0086 λ ∼ N (0.1, 0.082 )

The calibration fit is presented in Figures 5, 6 and 7. We report an excellent calibration fit
for all considered option expiries, varying from 1y to 20y. We have used two parameters in all
the calibration cases, just as in the HW model. The results confirm that the RAnD method has
great potential for improving existing pricing methods, even with the same number of degrees
of freedom.
17
 Calibration results for: 1Y-1Y Calibration results for: 2Y-1Y
28 30
Market Market
Hull-White Hull-White
randomized Hull-White 28 randomized Hull-White
26
ATM ATM

26
Implied Volatility [%] 24

Implied Volatility [%]

24

22

22

20
20

18
18

16 16
0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055
strike, K strike, K

Figure 5: Calibration results of the HW and the rHW models. The market implied
volatilities for swaptions were obtained on 18/08/2022 for the USD market. Option
expiry: T = 1y and T = 2y and the implied volatility shift: s = 1%. Calibrated
parameters are presented in Table 1.

Calibration results for: 5Y-1Y Calibration results for: 10Y-1Y

28 26
Market Market
26 Hull-White 24 Hull-White
randomized Hull-White randomized Hull-White
ATM ATM
22
24

20
Implied Volatility [%]

Implied Volatility [%]

22
18
20
16
18
14

16
12

14 10

12 8
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0 0.02 0.04 0.06 0.08 0.1 0.12
strike, K strike, K

Figure 6: Calibration results of the HW and the rHW models. The market implied
volatilities for swaptions were obtained on 18/08/2022 for the USD market. Option
expiry: T = 5y and T = 10y and the implied volatility shift: s = 1%. Calibrated
parameters are presented in Table 1.

Calibration results for: 15Y-1Y Calibration results for: 20Y-1Y

30 30
Market Market
Hull-White Hull-White
randomized Hull-White randomized Hull-White
25 ATM 25 ATM
Implied Volatility [%]

Implied Volatility [%]

20 20

15 15

10 10

5 5
0 0.05 0.1 0.15 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
strike, K strike, K

Figure 7: Calibration results of the HW and the rHW models. The market implied
volatilities for swaptions were obtained on 18/08/2022 for the USD market. Option
expiry: T = 15y and T = 20y and the implied volatility shift: s = 1%. Calibrated
parameters are presented in Table 1.

18
4.4. Pricing under bivariate distributions for the model parameters
This section extends the rHW model and considers a bivariate distribution for both model
parameters. Under a bivariate distribution Θ = [ϑ1 , ϑ2 ] with ζ(ϑ1 ) = {ω1,n , θ1,n }N
n=1 and
conditioned on ζ(ϑ2 |ϑ1 ) = {ω2,m , θ2,m }M
m=1 , the randomized prices are:

N
X M
X
VrHW (t, r(t); ϑ1 , ϑ2 ) = ωn ωm VHW (t, r(t); ηn , λm ) + bN,M , (4.3)
n=1 m=1

where N and M indicate the number of expansion terms for ϑ1 and ϑ2 |ϑ1 , respectively, ϑ2 |ϑ1
indicates a conditional random variable, bN,M is the corresponding aggregated error. The
remaining specification follows Theorem 2.1.
As an example, let us consider a bivariate normal distribution for the pair (ϑ1 , ϑ2 ) with the
corresponding realizations (η, λ) for which we have:
 
σλ 2 2
ϑ2 |ϑ1 = ηn ∼ N µλ + ρ(ηn − µη ), (1 − ρ )σλ , (4.4)
ση

where ρ is the correlation coefficient between ϑ1 and ϑ2 . In Figure 8, we illustrate the impact
of the correlation coefficient ρ on the swaption implied volatilities. We report that similar to,
e.g., the Heston model, the correlation controls the implied volatility skew. Furthermore, we
observe that a higher positive correlation generates more skew, while a negative correlation
generates more volatility curvature (smile).

Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
12 18

17
11
16

15
10
Implied Volatility, [%]

Implied Volatility, [%]

14

9 13

12
8
11

10
7
9

6 8
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)

Figure 8: Impact of randomized volatility parameters η and λ on implied volatilities

driven by bivariate normal distribution. Results are presented for varying correlation,
ρ. Left: η ∼ N (0.008, 0.0022 ) and λ ∼ N (0.5, 0.052 ). Right: η ∼ N (0.01, 0.0022 ) and
λ ∼ N (0.5, 0.22 ).

Remark 4.1 (Volatility term structure and feasible strategy for model calibration). Com-
monly, the volatility-parameter η is piece-wise constant, so the ATM volatilities are well cal-
ibrated. This strategy will also work with the randomized mean-reversion parameter, i.e., the
mean-reversion parameter can be used for smile/skew calibration, while the volatility parameter,
η, will ensure a proper fit of the ATM level.

4.5. Convergence results

As presented in Theorem 2.1, applying the RAnD method produces a quadrature error. In
this section, we analyze the convergence of the error depending on the number of expansion
terms, N .
The convergence speed depends on the distribution and its parameters, as presented in
Figure 9. We report an excellent convergence factor; however, as expected, it depends on the
variance of the randomizing random variable. Both cases will achieve satisfactory results for
already N = 5. Randomization using different random variables showed equivalent patterns of
convergence.
19
 -5 -5

-10
-10

-15
-15

-20

-20

-25

-25
-30

-30
-35

-35 -40
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 11
N N

Figure 9: Convergence results for randomization of volatility parameter, and mean re-
version parameter η and λ under the HW model respectively. The base parameters in
the experiment were η = 0.00625 and λ = 0.002.

5. Conclusion

In this paper, we have applied the randomization technique to enhance the flexibility of
short-rate models for interest rates. We have shown that the model parameters driven by a
random variable, instead of being deterministic, facilitate a practical extension of standard,
well-popularized models. In addition, this article points out how the normal mixture (a sum
of the Hull-White PDFs) can be expressed as a one-dimensional diffusion process with a local
volatility function.
We have illustrated that, for the randomized Hull-White model (rHW), one can utilize
the available closed-form pricing equations and benefit from flexibility in controlling implied
volatilities. In particular, we have shown that the rHW model results in almost perfect swaption
calibration.
Finally, the RAnD method is generic and is not limited to any particular modelling choice;
therefore, it opens many possibilities for improving existing pricing frameworks.

References
[1] J. Andreasen. Stochastic volatility interest rate models. Encyclopedia of Quantitative Finance, 2010.
[2] D. Brigo. The general mixture-diffusion SDE and its relationship with an uncertain-volatility option model
with volatility-asset decorrelation. arXiv preprint arXiv:0812.4052, 2008.
[3] D. Brigo and F. Mercurio. Lognormal-mixture dynamics and calibration to market volatility smiles.
International Journal of Theoretical and Applied Finance, 5(04):427–446, 2002.
[4] D. Brigo and F. Mercurio. Interest rate models- theory and practice: with smile, inflation and credit.
Springer Finance, 2nd edition, 2007.
[5] R. Bulirsch and J. Stoer. Introduction to numerical analysis, volume 3. Springer, 2002.
[6] J. Casassus, P. Collin-Dufresne, and B. Goldstein. Unspanned stochastic volatility and fixed income
derivatives pricing. Journal of Banking & Finance, 29(11):2723–2749, 2005.
[7] D. Duffie, J. Pan, and K. Singleton. Transform analysis and asset pricing for affine jump-diffusions.
Econometrica, 68:1343–1376, 2000.
[8] H. G. Fong and O. A. Vašiček. Fixed-income volatility management. Journal of Portfolio Management,
17(4):41, 1991.
[9] D. Ga̧tarek and J. Jablecki. A nonparametric local volatility model for swaptions smile. Journal of
Computational Finance, 21(5):35–62, 2018.
[10] G. H. Golub and J. H. Welsch. Calculation of Gauss quadrature rules. Mathematics of Computation,
23(106):221–230, 1969.
[11] L. A. Grzelak. On randomization of affine diffusion processes with application to pricing of options on
VIX and S&P 500. arXiv:2208.12518, 2022.
[12] L. A. Grzelak, J. Witteveen, M. Suarez-Taboada, and C. W. Oosterlee. The stochastic collocation Monte
Carlo sampler: highly efficient sampling from expensive distributions. Quantitative Finance, 19(2):339–
356, 2019.
[13] D. Heath, R. Jarrow, and A. Morton. Bond pricing and the term structure of interest rates: A new
methodology for contingent claims valuation. Econometrica, 1(60):77–105, 1992.
[14] J. Hull and A. White. Pricing interest-rate derivative securities. The Review of Financial Studies, 3:573–
592, 1990.
[15] F. Jamshidian. An exact bond option formula. The Journal of Finance, 44:205–209, 1989.

20
[16] C. W. Oosterlee and L. A. Grzelak. Mathematical modeling and computation in finance. World Scientific,
2019.
[17] L. C. G. Rogers and D. Williams. Diffusions, Markov processes and martingales, volume 2 of Cambridge
Mathematical Library. Cambridge University Press, 2 edition, 2000.
[18] O. Vašiček. An equilibrium characterization of the term structure. Journal of Financial Economics,
5:177–188, 1977.

Appendix A. Moments and optimal quadrature pairs, {ωn , θn }N

n=1

Appendix A.1. Moments and optimal quadrature pairs, {ωn , θn }N

n=1

Table A.2: Selected distributions for the stochastic parameters. For the normal random variable
with some mean, µ, and variance, σ 2 , it is sufficient to consider standard normal distribution,
N (0, 1), and properly scale the θn points, obtained from Appendix A.2.

name raw moment domain

n+1
−ân+1
ϑ ∼ U ([â, b̂]) E[ϑn ]= b̂ [â, b̂]
(n+1)(b̂−â)
ϑ ∼ exp(â) E[ϑn ] = ân!n R+
ϑ ∼ N (0, 1) E[ϑn ]
= (n − 1)!! if n even; 0 otherwise R
ϑ ∼ Γ(â, b̂) E[ϑn ] = b̂n Γ(n + â)/Γ(b̂) R+
Pn−1 (n−1)!2j−1
ϑ ∼ χ2 (â, b̂) E[ϑn ] = 2n−1 (n − 1)!(â + nb̂) + j=1 (n−j)!
(â + j b̂)E[ϑn−j ] R+ ∪ {0}

Appendix A.2. Optimal quadrature pairs, {ωn , θn }N n=1 , based on moments

Let {pi }N
i=0 with deg(p i ) = i be a sequence of orthogonal polynomials in L2 , with respect
to PDF, fϑ (ϑ), of ϑ, then the following holds,
Z
pi (x)pj (x)fϑ (x)dx = δi,j E p2i (ϑ) , i, j = 0, . . . , N,
 
E [pi (ϑ)pj (ϑ)] = (A.1)
R

with R the support of ϑ, δi,j the Kronecker delta. Orthogonal polynomials {pi }N
i=0 satisfy the
following recursion relation:

pi+1 (x) = (x − αi )pi (x) − βi pi−1 , i = 0, . . . , N − 1,

p−1 (x) ≡ 0, p0 (x) ≡ 1,

where αi and βi are determined in terms of the moments of a random variable ϑ. Consider
the monomials mi (ϑ) = ϑi , and define µi,j as µi,j = E [mi (ϑ)mj (ϑ)], From all moments µi,j ,
we construct the Gram matrix M = {µi,j }N i,j=0 , which is symmetric and contains all moments
1 2N
{1, E[ϑ ], . . . , E[ϑ ]}. Since matrix M is, by definition, positive definite [10], we decompose
M = RT R, by the Cholesky decomposition of M.
The following step relates the Cholesky upper-triangular matrix R to the orthogonal poly-
nomials. This relationship has been established in [10] and is given by,
 2
rj,j+1 rj−1,j rj+1,j+1
αj = − , j = 1, . . . , N, and βj = , j = 1, . . . , N − 1, (A.2)
rj,j rj−1,j−1 rj,j

with r0,0 = 1 and r0,1 = 0 and where ri,j is the (i, j)-th element of matrix R. Now, we
determine the symmetric tridiagonal matrix, J,
 √ 
√α1 β1 √0 0 0
 β1 α2 β2 0 0 
 
. . .
J :=  .. .. ..  ∈ RN ×N , (A.3)
 
 p p 
 0 0 βN −2 p αN −1 βN −1 
0 0 0 βN −1 αN

with spectral factorization:

J = W ΛW T , Λ = diag[λ∗1 , λ∗2 , . . . λ∗N ], W W T = I,

21
Due to the positivity of the off-diagonal entries, the eigenvalues λ∗j are distinct, and all W ’s
first row entries are non-vanishing. Moreover, it is well known [10] that the nodes and weights,
∗
{xi , wi }N T 2
i=1 , of the Gauss rule are, for i = 1, . . . , N , given by xi = λi and wi = (1 W i ) ,
1
where j is the ith axis vector .

1 The accompanying Python and MATLAB codes can be found at https://github.com/LechGrzelak/

Randomization.

22